Diagnosis of Significant Liver Fibrosis by Using a DCNN Model With Fusion of Features From US B-Mode Image and Nakagami Parametric Map: An Animal Study

Abstract

The assessment of liver fibrosis is usually required in the diagnostic and treatment procedures for chronic liver disease. In this paper, we propose a deep convolutional neural network (DCNN) with multi-feature fusion (DCNN-MFF) to extract and integrate features from ultrasound (US) B-mode image and Nakagami parametric map for significant liver fibrosis recognition. The DCNN-MFF model mainly consists of three branches, among which the first two branches are used for extracting deep features respectively from the US B-mode image and the Nakagami parametric map, while the remaining branch is designed for extracting quantitative US features from the Nakagami parametric map. At the backend of the DCNN model, the extracted deep and quantitative features are fused together for final decision making. The performance of the DCNN-MFF model was evaluated on an animal dataset collected from 84 rats with 168 liver lobes under various fibrosis stages. Across five-fold cross-validation, the accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) achieved by the proposed DCNN-MFF model for significant liver fibrosis recognition were respectively 0.827 (95% confidence interval [CI] 0.762–0.881), 0.821 (95% CI 0.729–0.892), 0.836 (95% CI 0.731–0.912), 0.869 (95% CI 0.809–0.916), which are significantly better than those provided by the comparative methods.

Keywords

• Liver fibrosis
• ultrasound
• deep learning
• Nakagami imaging
• feature fusion